Retina image segmentation method, system and device based on gated space convolution attention network, and storage medium

Abstract

The invention discloses a retinal image segmentation method, system and device based on a gated space convolution attention network, and a storage medium, and relates to the field of medical image processing, and the method comprises the steps: obtaining and inputting three-dimensional OCT and OCTA image data; dynamically adaptive compression (AUP) of the input data to efficiently preserve valid information; performing cross-modal difference complementary feature fusion (DCF) on the OCT and the OCTA; performing multi-scale channel attention fusion (MSCAF) on different scale features, highlighting a blood vessel related region and strengthening structural response; the fused features are subjected to gated space convolution attention network (GSCA-Net) segmentation, and high-precision segmentation of blood vessel crossing and complex boundary regions is achieved; and outputting a multi-task segmentation result. Through effective combination of the AUP, the DCF, the MSCAF and the improved GSCA-Net, efficient and accurate segmentation of capillary, artery, vein and FAZ regions is realized, the segmentation precision and robustness are significantly improved, the calculation burden is reduced, and the clinical practicability is improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and in particular to a method, system, device, and storage medium for retinal image segmentation based on a gated spatial convolutional attention network. Background Technology

[0002] The morphological features of retinal vessels and the avascular region of the fovea are crucial for the diagnosis of various diseases. Optical coherence tomography (OCT) and vascular imaging provide high-resolution technical support for their observation, while image segmentation is the core step in quantifying these features. Traditional segmentation methods, such as threshold-based methods, are susceptible to noise interference, while filter-based methods are parameter-sensitive, computationally expensive, and lack robustness in complex pathological images. Deep learning has driven breakthroughs in retinal segmentation, but most methods have limitations: 2D to 2D methods struggle to capture the three-dimensional distribution features of blood vessels, and 3D to 2D methods are prone to high computational costs or spatial information loss, and they do not fully integrate multimodal information from OCT and OCT-TA. Therefore, there is an urgent need for a retinal image segmentation method that can efficiently fuse multimodal features while balancing dimensionality reduction efficiency and information preservation. Summary of the Invention

[0003] Therefore, the technical problem to be solved by the present invention is to solve the problems of low dimensionality reduction efficiency, low multimodal feature fusion efficiency and low image segmentation accuracy in retinal image processing.

[0004] The above-mentioned technical problems are solved by the following technical solution: The present invention proposes a retinal image segmentation method, system, device and storage medium based on a gated spatial convolutional attention network, which includes a retinal image segmentation method based on a gated spatial convolutional attention network, a retinal image segmentation system based on a gated spatial convolutional attention network, a computer device and a computer-readable storage medium.

[0005] In a first aspect, the present invention proposes a retinal image segmentation method based on a gated spatial convolutional attention network, comprising the following steps: S1. Data Acquisition and Input Steps: Acquire 3D OCT and OCTA image data, input the original OCT and OCTA 3D data, and ensure the consistency and validity of the input data; S2. Adaptive Compression (AUP) Step: Dynamic AUP is performed on 3D OCT and OCT data to efficiently retain effective information in different depth ranges, thereby improving spatial information utilization and projection efficiency; S3. Differential Complementary Feature Fusion (DCF) Step: Extract and fuse the differential and complementary features of OCT and OCTTA to achieve effective integration of cross-modal information; S4. Multi-scale Channel Attention Fusion (MSCAF) step: After uniformly mapping features at different scales, cross-scale gating interaction is performed to highlight vascular-related regions and enhance structural responses; S5. Gated Spatial Convolutional Attention Network (GSCA-Net) Segmentation Steps: Utilizing the fused features, combined with gated spatial convolution (GSC) and channel attention mechanism (CAM), high-precision segmentation of blood vessel intersections and complex boundary regions is achieved; S6. Output Steps: Output multi-task segmentation results, including capillary, artery, vein and FAZ regions.

[0006] In a preferred embodiment of the retinal image segmentation method based on gated spatial convolutional attention network described in this invention: the data acquisition is to obtain three-dimensional volume data through a public database (OCTA_500 dataset), the OCT data can reflect the layered structure information of the retina, and the OCTA data can reflect the information on vascular perfusion and blood flow distribution.

[0007] In a preferred embodiment of the retinal image segmentation method based on a gated spatial convolutional attention network described in this invention: the adaptive compression (AUP) is a dynamic adaptive slice compression, and the specific steps are as follows. In sparse pixel regions, a thicker pixel slice is used for compression; in pixel regions where the amount of pixel information gradually increases, a thinner pixel slice with a thickness less than the thicker one is used for compression; and in the pixel core region where the information is most dense, an extremely thin pixel slice with a thickness less than the thinner one is used for compression. Each slice is compressed into a two-dimensional planar feature map through unidirectional pooling, and then stitched together along the depth direction to form a reduced three-dimensional feature map.

[0008] In a preferred embodiment of the retinal image segmentation method based on a gated spatial convolutional attention network described in this invention: the differential complementary feature fusion (DCF) includes complementary information fusion and differential feature extraction, and the specific steps are as follows. Complementary information fusion: Element-wise summation is performed on OCT and OCT features. Shared semantic information is extracted through a complementary feature fusion processor composed of convolution, normalization, and activation functions. Furthermore, an attention mechanism that jointly models channel and spatial features is used to enhance cross-modal complementarity. The specific calculation formula is as follows: in, F oct This represents the dimensionality-reduced OCT 3D feature map data. F octa This represents the 3D feature map data of OCTA after dimensionality reduction.f c This represents the result of element-wise summation of the features of OCT and OCTA. Relu(BN(Conv(·))) This represents a complementary feature fusion mechanism consisting of convolution, normalization, and activation functions. CA(·)×SA(·) This represents the attention mechanism computation for joint modeling of channels and space. F C This represents the result of complementary information fusion; Differential Feature Extraction: Element-wise absolute difference operations are used to obtain inter-modal differential features, which are then input together with the concatenated original features into a differential feature extractor consisting of convolution, normalization, and activation functions to obtain structured differential information. This differential information is weighted by a convolutional module attention mechanism combining channel and spatial attention, and then fused with the original features through residual connections. This process preserves the original semantic information while highlighting modal differences. The specific calculation formula is as follows: in, F oct This represents the dimensionality-reduced OCT 3D feature map data. F octa This represents the 3D feature map data of OCTA after dimensionality reduction. f d Indicates the differences between modes. cat(·) This indicates a concatenation operation. dim Representing dimension, f context This represents the original features after splicing. Relu(BN(Conv(·))) This represents a differential feature extractor composed of convolution, normalization, and activation functions. f cd Represents structured difference information. CBAM(·) This represents the attention mechanism operation of a convolutional module that combines channel and spatial attention mechanisms. F D This represents the result of fusing structured difference information and original features.

[0009] In a preferred embodiment of the retinal image segmentation method based on gated spatial convolutional attention networks described in this invention: the multi-scale channel attention fusion (MSCAF) comprises the following specific steps. By using a dynamic channel compressor consisting of 1×1 convolution, normalization, and activation function Gelu, input features at different levels are compressed and uniformly mapped to a baseline number of channels, reducing computational burden and ensuring feature consistency. By utilizing a cross-scale interactive gating mechanism, the contextual relationships of multi-scale features are captured through 3×3 convolution, and a gating signal is generated using the Sigmoid function to achieve dynamic weighting of features at different scales. By combining depthwise separable convolution and channel and spatial attention mechanisms, local spatial features of blood vessels are extracted, key semantic channels are screened, and spatial weights of blood vessel regions are enhanced. Output channels are adjusted using 1×1 convolutions, and residual connections are used to maintain training stability and feature consistency.

[0010] In a preferred embodiment of the retinal image segmentation method based on a gated spatial convolutional attention network (GSCA-Net) of the present invention: the gated spatial convolutional attention network (GSCA-Net) is an improved U-Net architecture employing a symmetric encoder and decoder, as detailed below. The encoder stacks triple convolutional units layer by layer, gradually expanding the receptive field from 3×3 to 7×7 to capture multi-scale information from local texture to global semantics. It also combines group normalization and the activation function PReLU to improve feature representation ability and training stability. The decoder restores spatial resolution through progressive upsampling and convolution, ensuring that the segmentation results have clear structure and accurate boundaries.

[0011] In a preferred embodiment of the retinal image segmentation method based on gated spatial convolutional attention network described in this invention: the encoder also introduces a GSC mechanism, which consists of standard convolutional branches and gated branches, and generates a gated weight map to dynamically adjust the features; The skip connection part introduces CAM, which generates a channel saliency map through global pooling and convolution, highlighting key channels and suppressing redundant channels.

[0012] Secondly, this invention proposes a retinal image segmentation system based on a gated spatial convolutional attention network, comprising: The data acquisition and input module is used to acquire 3D OCT and OCTA image data, input the original OCT and OCTA 3D data, and ensure the consistency and validity of the input data. The Adaptive Compression (AUP) module is used to dynamically perform AUP on 3D OCT and OCT data, efficiently retaining effective information in different depth ranges to improve spatial information utilization and projection efficiency. The Differential Complementary Feature Fusion (DCF) module is used to extract and fuse the differential and complementary features of OCT and OCTTA to achieve effective integration of cross-modal information. The Multi-Scale Channel Attention Fusion (MSCAF) module is used to perform cross-scale gating interaction after uniformly mapping features at different scales, highlighting vascular-related regions and enhancing structural responses; The Gated Spatial Convolutional Attention Network (GSCA-Net) segmentation module includes a GSC sub-module and a CAM sub-module. The GSCA-Net segmentation module is used to utilize the fused features, combining GSC and CAM, to achieve high-precision segmentation of blood vessel crossings and complex boundary regions. The results output module is used to output multi-task segmentation results, including capillary, artery, vein and FAZ regions.

[0013] Thirdly, the present invention proposes a computer device, including a memory, a processor, and computer-executable instructions stored in the memory, wherein the processor is used to execute the computer-executable instructions to implement the steps of the retinal image segmentation method based on a gated spatial convolutional attention network.

[0014] Fourthly, the present invention proposes a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the retinal image segmentation method based on a gated spatial convolutional attention network.

[0015] The beneficial effects of this invention are as follows: AUP compression of 3D OCT and OCTA volumetric data preserves vascular structural details, effectively reducing computational load; DCF utilizes feature algebra operations and cross-modal attention mechanisms to achieve differential extraction and complementary enhancement of structural information and blood flow features; MSCAF combines dynamic channel compression and cross-scale gating mechanisms to highlight vascular structural responses and suppress background noise; the improved GSCA-Net, through GSC and CAM, achieves excellent segmentation accuracy while maintaining low memory usage and inference time, especially in arteriovenous differentiation and identification of fine structures at vascular terminals. More importantly, this invention possesses good clinical adaptability and scalability, not only applicable to the auxiliary diagnosis and follow-up of diseases such as diabetic retinopathy and age-related macular degeneration, but also quickly transferable to other ocular structure segmentation tasks, such as optic disc and retinal nerve fiber layer segmentation; it can also be extended to fundus lesion detection tasks, such as the identification of microaneurysms and exudates in diabetic retinopathy. Therefore, this invention surpasses existing technologies in accuracy, efficiency, and scalability, possessing significant practical value. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention. Wherein: Figure 1 A flowchart of a retinal image segmentation method based on a gated spatial convolutional attention network is shown; Figure 2A schematic diagram of a retinal image segmentation system based on a gated spatial convolutional attention network is shown. Figure 3 A schematic diagram of the 640-pixel depth adaptive compression (AUP) process is shown; Figure 4 A schematic diagram of the differential complementary feature fusion (DCF) process is shown; Figure 5 A schematic diagram of the multi-scale channel attention fusion (MSCAF) process is shown; Figure 6 A schematic diagram of the gated spatial convolution (GSC) process is shown; Figure 7 A comparison chart showing the visualization effects of the segmentation results is provided; Figure 8 A comparison chart of ablation experiment results is shown. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0018] The terminology used in this invention is that which is currently widely used in the art in consideration of the function of the invention; however, these terms may vary according to the intent of those skilled in the art, precedent, or new technology in the art. Furthermore, specific terms may be chosen by the applicant, and in such cases, their detailed meanings will be described in the detailed description of the invention. Therefore, the terms used in this specification should not be construed as simple names, but rather based on their meanings and the overall description of the invention.

[0019] Reference Figure 1 , Figure 3 , Figure 4 , Figure 5 and Figure 6 This embodiment provides a retinal image segmentation method based on a gated spatial convolutional attention network, including the following steps: S1. Data Acquisition and Input Steps: Acquire 3D OCT and OCTA image data, input the original OCT and OCTA 3D data, and ensure the consistency and validity of the input data.

[0020] Specifically, the data acquisition is achieved by obtaining three-dimensional volume data from a public database (OCTA_500 dataset). The OCT data can reflect the layered structure information of the retina, while the OCTA data can reflect the information on vascular perfusion and blood flow distribution. The two have natural differences and complementarity. To ensure the integrity of the effective information in the data, the original OCT and OCTA three-dimensional data are input.

[0021] S2. Adaptive Compression (AUP) Step: Dynamic AUP is performed on 3D OCT and OCT data to efficiently retain effective information in different depth ranges, thereby improving spatial information utilization and projection efficiency.

[0022] S2 specifically includes the following steps. S2-1. In sparse pixel segments, a larger thickness pixel slice is used for compression; in pixel segments where the amount of pixel information gradually increases, a thinner thickness pixel slice is used for compression; and in the pixel core segment where the information is most dense, an extremely thin thickness pixel slice is used for compression. Specifically, refer to Figure 3 Taking a depth of 640 pixels as an example: in the sparse information ranges of 0 to 200 pixels and 500 to 640 pixels, pixel slices with larger thicknesses (20 pixels and 14 pixels) are used for compression; in the ranges of 200 to 300 pixels and 400 to 500 pixels where the amount of information gradually increases, pixel slices with a relatively thin thickness of 10 pixels are used for compression; in the core range of 300 to 400 pixels where the information is most dense, pixel slices with an extremely thin thickness of 5 pixels are used for compression.

[0023] S2-2. Each slice is compressed into a two-dimensional planar feature map by unidirectional pooling, and then stitched together along the depth direction to form a dimension-reduced three-dimensional feature map.

[0024] S3. Differential Complementary Feature Fusion (DCF) Step: Extract and fuse the differential and complementary features of OCT and OCTTA to achieve effective integration of cross-modal information.

[0025] Reference Figure 4 S3 includes complementary information fusion and differential feature extraction, and the specific steps are as follows: Complementary information fusion: Element-wise summation is performed on OCT and OCT features. Shared semantic information is extracted through a complementary feature fusion processor composed of convolution, normalization, and activation functions. Furthermore, an attention mechanism that jointly models channel and spatial features is used to enhance cross-modal complementarity. The specific calculation formula is as follows: in, F oct This represents the dimensionality-reduced OCT 3D feature map data. F octa This represents the 3D feature map data of OCTA after dimensionality reduction. f c This represents the result of element-wise summation of the features of OCT and OCTA. Relu(BN(Conv(·))) This represents a complementary feature fusion mechanism consisting of convolution, normalization, and activation functions. CA(·)×SA(·) This represents the attention mechanism computation for joint modeling of channels and space. F C This represents the result of complementary information fusion; Differential Feature Extraction: Element-wise absolute difference operations are used to obtain inter-modal differential features, which are then input together with the concatenated original features into a differential feature extractor consisting of convolution, normalization, and activation functions to obtain structured differential information. This differential information is weighted by a convolutional module attention mechanism combining channel and spatial attention, and then fused with the original features through residual connections. This process preserves the original semantic information while highlighting modal differences. The specific calculation formula is as follows: in, F oct This represents the dimensionality-reduced OCT 3D feature map data. F octa This represents the 3D feature map data of OCTA after dimensionality reduction. f d Indicates the differences between modes. cat(·) This indicates a concatenation operation. dim Representing dimension, f context This represents the original features after splicing. Relu(BN(Conv(·))) This represents a differential feature extractor composed of convolution, normalization, and activation functions. f cd Represents structured difference information. CBAM(·) This represents the attention mechanism operation of a convolutional module that combines channel and spatial attention mechanisms. F D This represents the result of fusing structured difference information and original features.

[0026] S4. Multi-scale Channel Attention Fusion (MSCAF) step: After uniformly mapping features at different scales, cross-scale gating interaction is performed to highlight vascular-related regions and enhance structural responses.

[0027] Reference Figure 5 S4 specifically includes the following steps: S4-1. By using a dynamic channel compressor consisting of 1×1 convolution, normalization and activation function Gelu, the input features at different levels are compressed and uniformly mapped to the baseline number of channels, reducing the computational burden and ensuring feature consistency. S4-2. By utilizing a cross-scale interactive gating mechanism, the contextual relationships of multi-scale features are captured through 3×3 convolution, and the gating signal is generated using the Sigmoid function to achieve dynamic weighting of features at different scales. S4-3. Combining depthwise separable convolution with channel and spatial attention mechanisms, we extract local spatial features of blood vessels, filter key semantic channels, and enhance the spatial weight of blood vessel regions. S4-4. Output channels are adjusted using 1×1 convolutions, and residual connections are used to maintain training stability and feature consistency.

[0028] S5. Gated Spatial Convolutional Attention Network (GSCA-Net) segmentation steps: Utilize the fused features, combine GSC and CAM to achieve high-precision segmentation of blood vessel intersections and complex boundary regions.

[0029] Specifically, the gated spatial convolutional attention network (GSCA-Net) is an improved U-Net architecture that employs a symmetric encoder and decoder; The encoder stacks triple convolutional units layer by layer, gradually expanding the receptive field from 3×3 to 7×7 to capture multi-scale information from local texture to global semantics. It also combines group normalization and the activation function PReLU to improve feature representation ability and training stability. Reference Figure 6 Furthermore, a GSC mechanism is introduced into the encoder, which consists of standard convolutional branches and gated branches, generating a gated weight map to dynamically adjust the features; The skip connection part introduces CAM, which generates a channel saliency map through global pooling and convolution to highlight key channels and suppress redundant channels; The decoder restores spatial resolution through progressive upsampling and convolution, ensuring that the segmentation results have clear structure and accurate boundaries.

[0030] S6. Output Steps: Output multi-task segmentation results, including capillary, artery, vein and FAZ regions.

[0031] Reference Figure 7 and Figure 8 This embodiment also provides an application scenario for the retinal image segmentation method based on gated spatial convolutional attention network: segmentation of retinal tissue structures (capillaries, arteries, veins, FAZ), where AUP and DCF can also be extended to other multimodal medical image segmentation tasks.

[0032] from Figure 7As can be seen, compared with other segmentation methods, the method of the present invention outputs multi-task segmentation results including capillaries, arteries, veins and FAZ regions, and the final output segmentation results have clear boundaries and complete structure, which can meet the needs of clinical analysis and auxiliary diagnosis.

[0033] from Figure 8 As can be seen, the improved GSCA-Net demonstrates significant advantages in both subsets of the OCTA-500 dataset, surpassing the current state-of-the-art (SOTA) methods in average Intersection over Union (IoU), reaching 82.84% and 88.48% respectively; and achieving average Dice coefficients of 90.40% and 93.76% respectively. Specifically, on the OCTA_6mm and OCTA_3mm datasets, the IoU coefficients for capillaries, arteries, and veins using FAZ are 85.49%, 79.08%, 79.42%, 87.96%, 88.47%, 85.25%, 84.76%, and 95.91%, respectively, while the Dice coefficients are 92.47%, 87.96%, 88.32%, 92.78%, 93.41%, 91.16%, 91.22%, and 97.83%, respectively.

[0034] Reference Figure 2 This embodiment also provides a retinal image segmentation system based on a gated spatial convolutional attention network, including, The data acquisition and input module is used to acquire 3D OCT and OCTA image data, input the original OCT and OCTA 3D data, and ensure the consistency and validity of the input data. The Adaptive Compression (AUP) module is used to dynamically perform AUP on 3D OCT and OCT data, efficiently retaining effective information in different depth ranges to improve spatial information utilization and projection efficiency. The Differential Complementary Feature Fusion (DCF) module is used to extract and fuse the differential and complementary features of OCT and OCTTA to achieve effective integration of cross-modal information. The Multi-Scale Channel Attention Fusion (MSCAF) module is used to perform cross-scale gating interaction after uniformly mapping features at different scales, highlighting vascular-related regions and enhancing structural responses; The Gated Spatial Convolutional Attention Network (GSCA-Net) segmentation module includes a GSC sub-module and a CAM sub-module. The GSCA-Net segmentation module is used to utilize the fused features, combining GSC and CAM, to achieve high-precision segmentation of blood vessel crossings and complex boundary regions. The results output module is used to output multi-task segmentation results, including capillary, artery, vein and FAZ regions.

[0035] This embodiment also provides a computer device applicable to the retinal image segmentation method based on a gated spatial convolutional attention network, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method proposed in the above embodiment.

[0036] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0037] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the retinal image segmentation method based on a gated spatial convolutional attention network as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0038] In summary, this invention preserves vascular structural details while compressing 3D OCT and OCTA volumetric data using AUP, effectively reducing computational load. DCF utilizes feature algebra operations and cross-modal attention mechanisms to achieve differential extraction and complementary enhancement of structural information and blood flow features. MSCAF combines dynamic channel compression and cross-scale gating mechanisms to highlight vascular structural responses and suppress background noise. The improved GSCA-Net, through GSC and CAM, achieves excellent segmentation accuracy while maintaining low memory usage and inference time, particularly in arteriovenous differentiation and identification of fine structures at vascular terminals. More importantly, this invention possesses good clinical adaptability and scalability, applicable not only to the auxiliary diagnosis and follow-up of diseases such as diabetic retinopathy and age-related macular degeneration, but also quickly transferable to other ocular structure segmentation tasks, such as optic disc and retinal nerve fiber layer segmentation; it can also be extended to fundus lesion detection tasks, such as the identification of microaneurysms and exudates in diabetic retinopathy. Therefore, this invention surpasses existing technologies in accuracy, efficiency, and scalability, demonstrating significant practical value.

[0039] Finally, it should be noted that the methods, systems, devices, and storage media described in detail above are merely embodiments, and those skilled in the art can modify these embodiments in different ways as long as they do not depart from the scope of the present invention.

Claims

1. A retinal image segmentation method based on a gated spatial convolutional attention network, characterized in that: Includes the following steps, S1. Data Acquisition and Input Steps: Acquire 3D OCT and OCTA image data, input the original OCT and OCTA 3D data, and ensure the consistency and validity of the input data; S2. Adaptive Compression (AUP) Step: Dynamic AUP is performed on 3D OCT and OCT data to efficiently retain effective information in different depth ranges, thereby improving spatial information utilization and projection efficiency; S3. Differential Complementary Feature Fusion (DCF) Step: Extract and fuse the differential and complementary features of OCT and OCTTA to achieve effective integration of cross-modal information; S4. Multi-scale Channel Attention Fusion (MSCAF) step: After uniformly mapping features at different scales, cross-scale gating interaction is performed to highlight vascular-related regions and enhance structural responses; S5. Gated Spatial Convolutional Attention Network (GSCA-Net) Segmentation Steps: Utilizing the fused features, combined with gated spatial convolution (GSC) and channel attention mechanism (CAM), high-precision segmentation of blood vessel intersections and complex boundary regions is achieved; S6. Output Steps: Output multi-task segmentation results, including capillary, artery, vein and FAZ regions.

2. The retinal image segmentation method based on a gated spatial convolutional attention network according to claim 1, characterized in that: The data acquisition is achieved by obtaining three-dimensional volume data from a public database (OCTA_500 dataset). The OCT data can reflect the layered structure information of the retina, and the OCTA data can reflect the information on vascular perfusion and blood flow distribution.

3. The retinal image segmentation method based on a gated spatial convolutional attention network according to claim 2, characterized in that: The adaptive compression (AUP) is a dynamic adaptive slice compression, and the specific steps are as follows. In sparse pixel regions, a thicker pixel slice is used for compression; in pixel regions where the amount of pixel information gradually increases, a thinner pixel slice with a thickness less than the thicker one is used for compression; and in the pixel core region where the information is most dense, an extremely thin pixel slice with a thickness less than the thinner one is used for compression. Each slice is compressed into a two-dimensional planar feature map through unidirectional pooling, and then stitched together along the depth direction to form a reduced three-dimensional feature map.

4. The retinal image segmentation method based on a gated spatial convolutional attention network according to claim 3, characterized in that: The differential complementary feature fusion (DCF) includes complementary information fusion and differential feature extraction, and the specific steps are as follows: Complementary information fusion: Element-wise summation is performed on OCT and OCT features. Shared semantic information is extracted through a complementary feature fusion processor composed of convolution, normalization, and activation functions. Furthermore, an attention mechanism that jointly models channel and spatial features is used to enhance cross-modal complementarity. The specific calculation formula is as follows: in, F oct This represents the dimensionality-reduced OCT 3D feature map data. F octa This represents the 3D feature map data of OCTA after dimensionality reduction. f c This represents the result of element-wise summation of the features of OCT and OCTA. Relu(BN(Conv(·))) This represents a complementary feature fusion mechanism consisting of convolution, normalization, and activation functions. CA(·)×SA(·) This represents the attention mechanism computation for joint modeling of channels and space. F C This represents the result of complementary information fusion; Differential Feature Extraction: Element-wise absolute difference operations are used to obtain inter-modal differential features, which are then input together with the concatenated original features into a differential feature extractor consisting of convolution, normalization, and activation functions to obtain structured differential information. This differential information is weighted by a convolutional module attention mechanism combining channel and spatial attention, and then fused with the original features through residual connections. This process preserves the original semantic information while highlighting modal differences. The specific calculation formula is as follows: in, F oct This represents the dimensionality-reduced OCT 3D feature map data. F octa This represents the 3D feature map data of OCTA after dimensionality reduction. f d Indicates the differences between modes. cat(·) This indicates a concatenation operation. dim Representing dimension, f context This represents the original features after splicing. Relu(BN(Conv(·))) This represents a differential feature extractor composed of convolution, normalization, and activation functions. f cd Represents structured difference information. CBAM(·) This represents the attention mechanism operation of a convolutional module that combines channel and spatial attention mechanisms. F D This represents the result of fusing structured difference information and original features.

5. The retinal image segmentation method based on a gated spatial convolutional attention network according to claim 4, characterized in that: The specific steps of the Multi-Scale Channel Attention Fusion (MSCAF) are as follows. By using a dynamic channel compressor consisting of 1×1 convolution, normalization, and activation function Gelu, input features at different levels are compressed and uniformly mapped to a baseline number of channels, reducing computational burden and ensuring feature consistency. By utilizing a cross-scale interactive gating mechanism, the contextual relationships of multi-scale features are captured through 3×3 convolution, and a gating signal is generated using the Sigmoid function to achieve dynamic weighting of features at different scales. By combining depthwise separable convolution and channel and spatial attention mechanisms, local spatial features of blood vessels are extracted, key semantic channels are screened, and spatial weights of blood vessel regions are enhanced. Output channels are adjusted using 1×1 convolutions, and residual connections are used to maintain training stability and feature consistency.

6. The retinal image segmentation method based on a gated spatial convolutional attention network according to claim 5, characterized in that: The gated spatial convolutional attention network (GSCA-Net) is an improved U-Net architecture employing a symmetric encoder and decoder, as detailed below. The encoder stacks triple convolutional units layer by layer, gradually expanding the receptive field from 3×3 to 7×7 to capture multi-scale information from local texture to global semantics. It also combines group normalization and the activation function PReLU to improve feature representation ability and training stability. The decoder restores spatial resolution through progressive upsampling and convolution, ensuring that the segmentation results have clear structure and accurate boundaries.

7. The retinal image segmentation method based on a gated spatial convolutional attention network according to claim 6, characterized in that: The encoder also introduces a GSC mechanism, which consists of a standard convolutional branch and a gated branch, and generates a gated weight map to dynamically adjust the features. The skip connection part introduces CAM, which generates a channel saliency map through global pooling and convolution, highlighting key channels and suppressing redundant channels.

8. A retinal image segmentation system based on a gated spatial convolutional attention network, wherein the method described in any one of claims 1-7 is characterized in that: include, The data acquisition and input module is used to acquire 3D OCT and OCTA image data, input the original OCT and OCTA 3D data, and ensure the consistency and validity of the input data. The Adaptive Compression (AUP) module is used to dynamically perform AUP on 3D OCT and OCT data, efficiently retaining effective information in different depth ranges to improve spatial information utilization and projection efficiency. The Differential Complementary Feature Fusion (DCF) module is used to extract and fuse the differential and complementary features of OCT and OCTTA to achieve effective integration of cross-modal information. The Multi-Scale Channel Attention Fusion (MSCAF) module is used to perform cross-scale gating interaction after uniformly mapping features at different scales, highlighting vascular-related regions and enhancing structural responses; The Gated Spatial Convolutional Attention Network (GSCA-Net) segmentation module includes a GSC sub-module and a CAM sub-module. The GSCA-Net segmentation module is used to utilize the fused features, combining GSC and CAM, to achieve high-precision segmentation of blood vessel crossings and complex boundary regions. The results output module is used to output multi-task segmentation results, including capillary, artery, vein and FAZ regions.

9. A computer device, comprising a memory, a processor, and computer-executable instructions stored in the memory, characterized in that: The processor is configured to execute computer-executable instructions to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps of the method described in any one of claims 1-7.

Images